Fig. 1

LoGoFunc workflow and model architecture. a Pipeline for the collection of labeled pathogenic GOF and LOF variants. Related abstracts for high confidence pathogenic variants from the HGMD [15] were searched for nomenclature denoting gain or loss of function. b Dataset preparation and annotation. 1492 GOF, 13,524 LOF, and 13,361 neutral variants were obtained from the GOF/LOF database [14], HGMD, and gnomAD [17]. Using VEP [21] and other tools, variants were annotated with protein structural and functional features derived from AlphaFold2 [12] models or from sequence, with gene- and genomic-level features, variant-level features, and network-derived protein interaction features. The annotated data were split into training and test sets comprising 90% and 10% of the dataset respectively, stratified by variant label. c Model architecture and output. Variants are input to the model represented as an array of the 474 collected features. These features are encoded, imputed, and scaled prior to prediction. The model consists of an ensemble of 27 LightGBM [11] classifiers. A probability is output for each class, GOF, LOF, and neutral